彎道之影像偵測與自動駕駛之實現

Abstract

自動駕駛是ITS近幾年來主要的研究方向，現今也已發展出不同的方式來完成這項工作。本研究選擇以CCD為感測器來擷取外界環境，其中，對我們最主要的特徵就是車道線。在直線道路上，沿用IPM based之車道線偵測方法，找出車道線偵測視窗之定位，再由動態視窗進行對車道線追跡，此方式能兼具辨識率及即時性的優點，並估測出虛擬的車道中心線作為導航的依據。但在彎道上，存在紅色連續的右車道標記及不連續的左車道標記，造成原偵測方法因不連續及紅色標記而失敗，故在彎道上，為了突顯紅色標記，以色度座標轉換及條件式二值化取代原先的二值化步驟，使原偵測法則亦能有效偵測彎道上的右車道線，並做為導航的依據。 從車道線可進一步測出車輛在道路上的相關狀態: (1) 前視偏移距離，(2) 偏轉角度，及(3) 道路曲率。在直線道路上，使用P控制器對車輛進行控制，根據實驗結果顯示，當給予 =5~11時，能獲得穩定及準確的結果。但在彎道上，由於曲率的存在容易影響車子偏離車道，故除了P控制器外，本文提出多加入一以曲率為輸入之模糊推論器建立對彎道之偏移量補償器，以改善行駛彎道時偏離車道情形。綜合以上之對直線及彎道之控制補償設計，本研究整合影像與控制法則，完成能在校園環校道路上直線及彎道的自動駕駛。

This thesis applies a CCD sensor to detect the lane marks for the present autonomous driving system. As the road is straight, the IPM (inverse perspective mapping)-based detection method is adopted to locate the suitable windows position properly. Then, by using the on-line dynamic detection windows approach, the road lanes can be thus efficiently tracked and the estimated virtual center-lane can be further obtained for the present autonomous driving. However, as on the curved road, some discontinuous lane marks cause the developed lane detection method invalid. Thus, the image of the red lane marks alone the road side can be processed by using the color-space transform with the conditional threshold to result in more accurate detection results in the case of curved roads. The vehicle states can be further estimated from the processed image as (1) the look-ahead offset, (2) the yaw angle, and (3) the road curvature. In the case of a straight road, the P controller is suitable for the autonomous vehicle and experimental results indicate that Kp between 5 to 11 leads to satisfactory performance by concerning the system stability and control accuracy. On the curved road, a curved road offset compensator is processed by applying the fuzzy inference, with the input of curvature, to significantly improve vehicle driving performance for the present autonomous system.